Abstract
Abstract
Owing to their favorable weight-to-performance characteristics, composite sandwich panels are extensively employed across engineering sectors; however, internal defects often escape identification when relying on conventional, visually-driven inspection routines. This study proposes a data-driven framework combining modal shape, image processing, and a multi-head convolutional neural network (CNN) for automated damage identification, localization, and extent estimation. Damage is numerically modeled by introducing localized regions of reduced stiffness through the insertion of low-stiffness material in the core, skin, or interface layers, enabling realistic representation of structural degradation. The first five non-null vibration mode shapes are transformed into contour images and used as inputs to the proposed architecture. A multi-head CNN processes each vibration mode independently to enhance feature extraction and structural condition classification into four categories: undamaged, core damage, skin damage, and interface damage. Subsequently, a dedicated CNN predicts damage position and geometry based on two spatial coordinates and three shape parameters. Among the evaluated image processing techniques, the residual approach yielded the best performance, achieving F1-scores of 0.99 for core damage and 0.87 for both skin and interface damage. The framework also achieved high localization capability, with an average position error of approximately 4% and maximum relative errors below 5% of the plate dimensions; these localization and sizing results were obtained using the ground-truth damage class to route each sample to its corresponding regression network, and therefore do not reflect the potential propagation of classification errors through the complete classification-routing-regression pipeline. Compared with a previously developed baseline model, the proposed framework reduced the mean absolute error for most geometric parameters, with improvements exceeding 30% in several localization and sizing tasks. Despite challenges in angular orientation prediction, these results demonstrate the potential of the proposed methodology for structural health monitoring of composite sandwich structures.